MCP-Server-for-CFD
# CFD FastAPI Server
將 2D 翼型空氣動力分析包裝成一般 FastAPI HTTP API,讓前端、腳本或其他服務可以直接呼叫完整的「幾何 → 網格 → 求解 → 後處理 → 視覺化」流程。
```text
Client ──▶ FastAPI Server ──▶ aerosandbox / NeuralFoil / SU2
└─▶ Cp 分布 / 極曲線 / dashboard / artifact downloads
```
## 功能
| API | 功能 |
|---|---|
| `POST /api/geometry/airfoil` | 從 NACA 4/5 位數代碼生成翼型座標 (`.dat`) |
| `POST /api/meshes/2d` | 生成 2D 網格,支援 placeholder 與 SU2/Gmsh |
| `POST /api/solver/run` | 執行 NeuralFoil,或提交非同步 SU2 求解 |
| `GET /api/solver/results/{job_id}` | 查詢 job 狀態;完成後回傳 CL / CD / CM / Cpmin / 轉捩點 |
| `POST /api/workflow/airfoil` | 一次完成 geometry -> mesh -> solver;SU2 只提交背景求解 |
| `POST /api/visualizations` | 生成視覺化 artifact metadata |
| `GET /artifacts/{artifact_id}` | 在瀏覽器 inline 預覽視覺化圖檔 |
| `GET /artifacts/{artifact_id}/download` | 下載視覺化圖檔 |
| `/mcp/` | Streamable HTTP MCP endpoint,提供同一套 CFD workflow tools |
### 視覺化模式
| `plot_kind` | 內容 |
|---|---|
| `summary` | 翼型形狀 + 結果數字 |
| `cp` | 沿翼型表面真 Cp 分布 |
| `polar` | CL/CD vs alpha 雙軸極曲線 |
| `drag_polar` | CL vs CD 拖力極曲線 |
| `mesh` | 網格預覽圖 |
| `dashboard` | 綜合 dashboard 與網格預覽 |
| `mach` | 馬赫數分佈圖 (SU2) |
| `pressure` | 壓力分佈圖 (SU2) |
`POST /api/visualizations` 會回傳:
```json
{
"status": "success",
"summary": "...",
"artifact": {
"artifact_id": "vis_abc123",
"filename": "dashboard_vis_abc123.jpg",
"mimeType": "image/jpeg",
"size_bytes": 170868
}
}
```
圖檔不直接塞進 JSON body,而是落地成 artifact。用回傳的 `artifact_id` 組 `GET /artifacts/{artifact_id}` 在瀏覽器預覽,或組 `GET /artifacts/{artifact_id}/download` 下載。
HTTP API 只回傳資源 id,不回傳伺服器端檔案路徑。手動流程請依序傳遞 `geometry_id`、`mesh_id`、`job_id`、`artifact_id`。
SU2 求解是非同步工作。`POST /api/solver/run` 或 `POST /api/workflow/airfoil` 使用 `solver_backend="su2"` 時會回傳 `status: "submitted"` 與 `job_id`;之後用 `GET /api/solver/results/{job_id}` 輪詢,直到狀態變成 `converged` 或 `failed`。NeuralFoil surrogate 仍同步回傳結果。
## 快速開始
### 環境需求
- Python 3.11+
- `uv`
### 本機啟動
```bash
uv sync
uv run python -m cfd_server
```
啟動後可用:
- Swagger UI: `http://localhost:8765/docs`
- OpenAPI: `http://localhost:8765/openapi.json`
- Health check: `http://localhost:8765/healthz`
- MCP endpoint: `http://localhost:8765/mcp/`
### 執行測試
```bash
uv run --group dev pytest
```
### Docker
```bash
docker build -t cfd-server:su2-local .
docker run --rm -p 8765:8765 -v "$PWD/jobs:/app/jobs" cfd-server:su2-local
```
檢查 SU2 與 Gmsh:
```bash
docker run --rm cfd-server:su2-local SU2_CFD --help
docker run --rm cfd-server:su2-local gmsh --version
```
### Docker Compose
```bash
docker compose up --build cfd-server
```
## API 範例
### 建立翼型
```bash
curl -X POST http://localhost:8765/api/geometry/airfoil \
-H "content-type: application/json" \
-d '{
"naca_code": "0012",
"chord_length": 1.0,
"n_points_per_side": 180,
"normalize_geometry": true
}'
```
### 一鍵工作流
```bash
curl -X POST http://localhost:8765/api/workflow/airfoil \
-H "content-type: application/json" \
-d '{
"naca_code": "0012",
"velocity": 30.0,
"angle_of_attack": 5.0,
"solver_backend": "neuralfoil"
}'
```
SU2 一鍵工作流會先產生 SU2/Gmsh mesh,然後提交背景求解:
```bash
curl -X POST http://localhost:8765/api/workflow/airfoil \
-H "content-type: application/json" \
-d '{
"naca_code": "0012",
"velocity": 30.0,
"angle_of_attack": 5.0,
"mesh_format": "su2",
"solver_backend": "su2",
"max_iterations": 500
}'
```
查詢背景求解:
```bash
curl http://localhost:8765/api/solver/results/06936164e5a7
```
### 視覺化
```bash
curl -X POST http://localhost:8765/api/visualizations \
-H "content-type: application/json" \
-d '{
"job_id": "06936164e5a7",
"plot_kind": "dashboard",
"image_format": "jpeg"
}'
```
## MCP
這個服務現在同時提供 Streamable HTTP MCP,直接重用現有 service layer,不需要另外維護第二套 CFD 邏輯。
MCP tools:
- `generate_airfoil_geometry`
- `generate_2d_mesh`
- `run_cfd_solver`
- `check_solver_results`
- `run_airfoil_workflow`
- `visualize_cfd_results`
- `get_visualization_artifact`
本機啟動 HTTP + MCP:
```bash
uv run python -m cfd_server
```
用 MCP Inspector 連線:
```bash
npx -y @modelcontextprotocol/inspector
```
連到:
```text
http://localhost:8765/mcp/
```
如果你要用 stdio 模式直接跑 MCP server:
```bash
uv run python -m cfd_server.app.interfaces.mcp.server
```
## 測試
- `tests/conftest.py`: 共用 FastAPI `TestClient` 與 workflow fixture
- `tests/test_http_workflow.py`: 幾何、網格、求解、結果查詢、路由註冊
- `tests/test_async_su2_solver.py`: SU2 非同步提交與狀態轉換
- `tests/test_visualization_artifacts.py`: artifact metadata 與落地檔案
- `tests/test_artifact_download.py`: artifact HTTP download endpoint
- `tests/test_su2_config.py`: SU2 config 單元測試
## 專案結構
```text
cfd-server/
├── cfd_server/
│ ├── __main__.py
│ ├── app/
│ │ ├── interfaces/
│ │ │ └── http/ # FastAPI app, routers, request schemas
│ │ └── services/ # workflow / visualization service layer
│ ├── server.py # 相容 wrapper
│ ├── core/
│ ├── mesh/
│ ├── solvers/
│ ├── visualization/
│ └── models/
├── tests/
├── pyproject.toml
├── Dockerfile
└── README.md
```
TDQS
Scored across 7 tools
Each tool maps to a distinct stage of the CFD workflow: geometry, meshing, solver execution, status checking, combined workflow, visualization, and artifact retrieval. The only near-overlap is run_cfd_solver versus run_airfoil_workflow, but their descriptions clearly separate step-by-step execution from end-to-end orchestration.
All tool names follow a consistent verb_noun pattern with snake_case (generate_, run_, check_, visualize_, get_). This uniformity makes the toolset predictable and easy to navigate.
Seven tools is well-scoped for an airfoil CFD pipeline, covering geometry generation, meshing, solver execution, result checking, visualization, and artifact retrieval without redundancy. Each tool earns its place.
The set covers the entire airfoil simulation lifecycle from geometry generation through visualization and artifact access. The inclusion of a combined workflow tool fills a potential gap for users who want a single-call solution. No essential operation appears missing.